HOSCAR project home page

High performance cOmputing and SCientific dAta management
dRiven by highly demanding applications

Project objectives and organization

The general objective of the project is to setup a Brazil-France collaborative effort for taking full benefits of the processing capabilities of future high performance massively parallel architectures in the framework of very large-scale numerical simulations relevant to a selected set of computational physics applications. In order to do so, the project involves:

• Computer scientists whose research and development activities aim at exploiting the processing capabilities and raw performances of massively parallel systems with high-quality and low-cost software.
• Algorithmists who propose algorithms and data structures that contribute to such software in order to take benefit from all the parallelism levels with the main goal of optimal scaling on very large numbers of computing entities.
• Numerical mathematicians who are studying numerical schemes and scalable solvers for systems of partial differential equations modeling applications related to important scientific questions for the quality and the security of life in our society.
  
  

All together, these researchers will form a continuum of skills on enabling methodologies and technologies based on the complementary expertise of the Brazilian and French researchers and associated groups working on the computer science and numerical mathematics aspects of high performance computing. More precisely, the following topics should be addressed:

• Numerical schemes for differential equations tailored to modern high performance systems. This topic is concerned with the study of strategies for the numerical treatment of the differential equations underlying the target physical applications. This includes discretization methods, in particular high order, possibly adaptive, finite element and finite volume methods, and also multiscale and multiresolution methods. Also of concern are coupling strategies for multi-physics applications.
• Numerical linear algebra. This topic is concerned with the development of core numerical linear algebra kernels adapted to modern high performance computing systems, as well as with the design of scalable solvers for the large linear systems of algebraic equations resulting from the numerical treatment of differential equations by the numerical schemes studied in the previous topic. This includes the solution of sparse linear systems which is one of the most critical and intensive computational kernels in terms of memory and time requirements, and which is very often at the heart of a numerical simulation tool.
• Optimization of performances of numerical solvers. This topic is concerned with different aspects of the enhancement of the performances of high performance numerical simulation tools. This includes: pre- and post-processing tools (parallel dynamic graph (re)-partitioning and re-meshing, load-balancing, computational steering), runtime systems, and scheduling and resource management strategies.
• Programming models for petascale computing. This topic is concerned with the adaptation and harnessing of new massively parallel programming paradigms for numerical computing applications. This includes the study of strategies that allow a transparent and efficient combined exploitation of fine grain and coarse grain parallelisms.
• Processing of very large datasets. This topic deals with the management of very large datasets that are manipulated (accessed and produced) by data-centric scientific workflows in HPC environments. Addressing the very scale of the datasets requires new scalable parallel data management techniques as well as scalable data-aware scheduling strategies. Furthermore, getting data in and out HPC environments from the scientists’ own environment is a major challenge. Finally, it is important to provide support for data provenance, a key function that records critical metadata about experiments to help scientists understanding results or reusing some workflow parts.
  
  

The project involves computer scientists and numerical mathematicians divided in 3 fundamental research groups: (i) Numerical schemes for PDE models (Group 1), (ii) Scientific data management (Group 2), and (iii) High-performance software systems (Group 3).